refactor(plan): extract caching state out of Visitor into cachingPlannerState

ysmolski flagged that the caching feature added ~900 LOC of two new
responsibilities on top of the planner visitor: a parallel ProvidesData
tree for caching, and per-fetch caching configuration. Visitor was
already 1600+ LOC before; this PR added another 900+ on top.

Extracts those two responsibilities into a sub-struct on Visitor:

  type Visitor struct {
      ...
      caching *cachingPlannerState
  }

Sub-struct (Option B) chosen over a sibling visitor (Option A) because
the highest-risk part of this code is walker callback ordering. A
sibling visitor adds ordering risk with AllowVisitor, datasource
planner visitors, DefferOnEnterField, and the cost visitor. A sub-struct
preserves callback ordering exactly, with only Visitor registered on
the walker.

Moved to caching_planner_state.go (936 LOC):
- Fields: entityAnalyticsCache, requestScopedVisibleResponseKeys,
  requestScopedFetchAliases, plannerObjects, plannerCurrentFields,
  plannerResponsePaths, plannerEntityBoundaryPaths.
- Methods: 27 caching-specific methods including configureFetchCaching,
  trackFieldForPlanner, popFieldsForPlanner,
  configureSubscriptionEntityCachePopulation, entityCacheAnalytics,
  polymorphicEntityCacheAnalytics, configureMutationEntityImpact, etc.
- Helper: extractKeyFields.

Preserved on Visitor (load-bearing):
- fieldPlanners — cost visitor captures by pointer.
- plannerFields — fetch deps/reasons read it.
- fieldEnclosingTypeNames — caching state reads through parent.
- Public RequestScopedFetchAlias — kept for external API stability,
  delegates to v.caching.fetchAlias.

Net diff: visitor.go shrunk from 2646 to 1766 LOC (-880, -33%).
New caching_planner_state.go: 936 LOC.

Three test files updated for moved helpers (call-site rewrites only,
no logic change):
- visitor_path_normalization_test.go
- request_scoped_provides_data_test.go
- visitor_subscription_entity_population_test.go

Also adds a "Code Comment Conventions" section to CLAUDE.md banning
PR/issue/reviewer references in code comments. Two such references
that crept in during this refactor have been removed from
caching_planner_state.go and visitor_path_normalization_test.go.

Tests: full plan, resolve, execution suites green.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: jensneuse

CLAUDE.md

@@ -170,6 +170,30 @@ request token) to prevent production abuse. The gate is in `GraphQLHandler.cachi
 `router/core/graphql_handler.go`. Disabling L1 via these headers also disables @requestScoped
 coordinate L1 (since it shares the `EnableL1Cache` flag).
 
+## Code Comment Conventions
+
+**Never reference pull requests, issue numbers, review threads, or reviewer names in code comments.**
+
+Comments live in the codebase forever and outlive the workflow context they were written in.
+A `PR #1259` reference is meaningful for two weeks and noise for the next ten years.
+Reviewer attribution (`as requested in ysmolski's review`, `addresses SkArchon's comment`) belongs in commit messages and PR descriptions, never in source files.
+
+If a comment exists to explain a non-obvious behavior, explain the **behavior**, not the historical reason it was added.
+
+```go
+// CORRECT — explains the invariant
+// isEntityRootField previously compared a non-normalized current path against a
+// normalized boundary path. Without normalizing here first, queries that wrap the
+// boundary in `... on User { ... }` cause the prefix check to silently fail.
+
+// WRONG — references the PR / review / ticket where the fix was discussed
+// Regression guard for the A42 bug in PR #1259 raised by ysmolski:
+// isEntityRootField previously compared a non-normalized current path...
+```
+
+This applies to all code comments — production, tests, doc comments, file headers.
+Commit messages may reference PRs and reviewers; code may not.
+
 ## Testing Conventions
 
 **Before writing or modifying any test, read the package's `CLAUDE.md` if one exists.**